Nano Banana prompt languages: Korean and Japanese, per Google

Google lists 15 language codes for best Gemini image performance, and both Korean (ko-KR) and Japanese (ja-JP) are on the list, so a Korean or Japanese prompt is inside the range Google recommends. The list is in the limitations section of Google's Image generation with Gemini page, read 2026-10-02. On Sume the prompt is a plain string on POST /v1/images, so you send the text as written; the Sume Image API page documents no prompt-language setting.

Which languages does Google list?

The page says: for best performance, use the following languages: EN, ar-EG, de-DE, es-MX, fr-FR, hi-IN, id-ID, it-IT, ja-JP, ko-KR, pt-BR, ru-RU, ua-UA, vi-VN, zh-CN. It says "best performance", not "only these", so other languages are not ruled out by the text, but they are not in the recommended set either.

Do I set a language field on Sume?

No. The Sume Image API request table has model, prompt, n, resolution, aspect_ratio, quality, output_format and the other fields documented on the page, and none of them selects a prompt language. Write the prompt in Korean or Japanese and send it. The same Image API page says a parameter the selected model does not list is rejected with 400 unsupported_parameter, so do not invent a language field.

What about Korean text inside the picture?

The Google list is about the language of the prompt. It does not say how well Hangul or kana render inside an image, and this post does not claim it. Test with your own copy: put the exact string in quotes in the prompt, render one image, and read the result before you run a batch.

If you need the same poster in several languages, send one request per language with the translated string. That keeps a failed render to one retry, and failed generations are not billed on Sume.

Should I translate the prompt to English first?

Google's list does not ask you to. If a Korean prompt gives a result you are unhappy with, try the same prompt in English once and compare the two images. That is a test you run, not a rule from either vendor page, and it costs one extra image at the model's catalog price.